DeskMAS
Multi-agent collaboration system for desktop robots
About the Project
DeskMAS is a desktop multi-agent collaboration project built at the intersection of HCI and embodied AI. It started from an Overcooked-inspired software prototype and evolved into a platform that connects natural-language task input with real robot execution.
What I built
I designed the multi-agent architecture around a ManagerAgent plus multiple worker agents, using shared state and turn-based coordination to let an LLM break down tasks and make collaborative decisions. The project then moved onto an ESP32-based robotics platform, where I helped define a standardized serial protocol and JSON command structure so planning and execution could stay decoupled.
System design
On the robotics side, the system used a top-mounted 720P camera and ArUco markers for real-time localization, together with A* global planning and APF local obstacle avoidance for navigation. The goal was not only to make a single demo work, but to explore how a more general desktop multi-agent platform could be built for embodied collaboration experiments.
Key Features
- ManagerAgent and worker-agent collaboration architecture
- Shared state and turn-based coordination flow
- Natural-language task execution for desktop robots
- Standardized ESP32 serial protocol and JSON commands
- Real-time localization with ArUco markers
- A* planning with APF obstacle avoidance
Challenges & Solutions
- Turning software multi-agent logic into physical robot collaboration
- Decoupling high-level planning from low-level hardware execution
- Keeping localization and navigation stable in real time
- Balancing research flexibility with platform generality
Status
Research Prototype
Timeline
Jul 2025 - Sep 2025
Role
Multi-agent Research
Interested in this project?
I'd love to hear your thoughts or discuss potential collaborations.